A series for engineering leaders

The Agentic CTO

Your company bought the AI tools. Every engineer has a licence. And the delivery graphs look exactly like they did last year.

That’s not a tooling problem — it’s an operating model problem. Copilots make individuals faster. They don’t change what your org can do, because the process around them still assumes humans do every step. Six parts on rebuilding that process, written from inside an org that already works this way.

Written for CTOs, VPs of Engineering and technical founders running teams of roughly 5–100 engineers, somewhere between “we bought licences” and “agents ship unsupervised”.

1,000+pull requests
1full-time engineer
$1.40per merged PR, all-in

One production repo, one full-time engineer. How these numbers were counted →

Where to start: read Part 1 for the operating model. If you want the stack, the gates and the bill before the philosophy, start with Part 0. If you already run agents and want the objection answered first, start with Part 4.

Part 0 · The foundation

Build the harness first

Agent = model + harness. You rent the model; the harness is the part you own. Here is ours, product names included: the contract, the nine gates on main, the QA agent's cost curve, and the bill — about $1.40 per merged PR.

Read the article →

Part 3

Your agents react

Incidents, alerts and inbound bugs. The reaction ladder agents climb while you sleep, the weekend our pager stayed silent, and the backlog that triages itself.

Read the article →

Part 4

Securing the agentic org

Blast radius beats cleverness: a free-fire staging zone, read-versus-write production, a break-glass agent behind MFA — and why your ticket tracker is an attack surface.

Read the article →

Part 6

Your sceptics are right

The greybeards are correct about AI code quality and the worried are correct that the org changes. You don't win either group with hype — you win them with a harness and honest maths.

Read the article →

Score yourself first

Seven questions, one per part

The one that makes you uncomfortable is the part to start with.

  1. When did a human last read a merged PR line by line — and what did they catch that the gates didn't?Part 0
  2. If your best engineer were off for a month, what would actually slow down: delivery, or the meetings about delivery?Part 1
  3. What did your process learn from last week's mistake — and where is that written so the next agent reads it?Part 2
  4. Would your pager stay silent through a real outage? How do you know?Part 3
  5. What's the worst thing an agent holding your staging credentials could do before lunch?Part 4
  6. Who in your company could open a pull request tomorrow — and who decides whether it merges?Part 5
  7. Name the two people quietly deciding whether your AI rollout happens. Have you answered them?Part 6

What I actually do

An agentic engineering operating-model review

Bought the tools and the graphs didn’t move? Here is how I move them — over a handful of sessions with you and, where it helps, your leads:

  • Find where agents are creating activity without outcomes — and what that is costing you.
  • Map your SDLC and pick the first step you can safely delegate, with the guarantee it has to keep.
  • Define the human gates and the autonomy thresholds that let the rest run unattended.
  • Set the baseline delivery and quality metrics you’ll be judged on — before you scale anything.
  • Design the security model: environments, read-vs-write, break-glass, untrusted input.
  • Coach you through the adoption — the part that is about people, not tooling.

Book a first conversation

Who’s writing this

I’m Ben Jones. I was CTO of an engineering org of 50+ and scaled Nuri to 120 people. Today I’m co-founder and CPTO at Herita — the repo behind the numbers above — and I coach CTOs and VPs of Engineering through this shift.

Not sure which of these is your bottleneck?

That’s usually the first thing we work out together. Tell me what your delivery looks like now and I’ll tell you what I’d change first.

Get in touch

Or follow along on LinkedIn →